Already 8 months old when the scout caught it.
Intent engineering
What it is
Intent engineering is the practice of encoding what an autonomous agent should be optimizing for over time — objectives, outcomes, constraints, trade-off priorities (e.g. "trust over speed over cost"), decision authority, and stop rules — as a structured, machine-actionable specification, rather than leaving that judgment implicit in a prompt. Where a prompt tells a model what to do right now, an intent specification is meant to keep governing an agent's behavior after the prompt's explicit instructions run out, in situations nobody wrote a rule for.
Why it's emerging now
As agents are given longer-running, more autonomous tasks, teams have kept hitting the same failure: an agent behaves fine on the happy path but drifts once it hits a situation its prompt or context didn't anticipate, optimizing for whatever's easiest to measure instead of what the business actually wants. Prompt engineering addresses the immediate instruction and context engineering addresses what the agent knows and can access, but neither specifies what the agent should value when it has to make a judgment call. Intent engineering was named to fill that gap, first as a product-management framework in January 2026 and then picked up independently through the spring as teams operationalized agents in production enterprise settings.
How it relates to what you know
It sits alongside prompt engineering and context engineering on your ledger as a third, deeper layer: prompt engineering governs the instruction, context engineering governs what's visible and retrievable, and intent engineering governs the values and trade-offs behind decisions the agent makes on its own. It's a peer concept to agent behavioral contracts (both try to make agent behavior predictable outside the happy path) but comes from the product/enterprise-adoption side rather than the safety-research side, focusing on business outcomes and decision authority rather than formal contracts or verification.